Research map / Image generation / Fast diffusion sampling
DVD-Quant: Data-free Video Diffusion Transformers Quantization
Zhiteng Li, Hanxuan Li, Junyi Wu, Kai Liu, Haotong Qin, Linghe Kong, Guihai Chen, Yulun Zhang, Xiaokang Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use BGR together with ARQ at W4A4 (ablation) | DVD-Quant · Aesthetic Quality (W4A4) | Aesthetic Quality: ARQ only 53.95 -> 59.57 (adding BGR) | Improved |
| add ARQ to BGR at W4A4 (ablation) | DVD-Quant · Aesthetic Quality (W4A4) | Aesthetic Quality: 43.26 -> 59.57 | Improved |
| add BGR at W4A6 (ablation) | DVD-Quant · Imaging Quality (W4A6) | Imaging Quality: 58.68 -> 61.93 | Improved |
| add ARQ at W4A6 (ablation) | DVD-Quant · Imaging Quality (W4A6) | Imaging Quality: 57.72 -> 61.93 | Improved |
| add ARQ to BGR at W4A4 (ablation) | DVD-Quant · Imaging Quality (W4A4) | Imaging Quality: 58.31 -> 58.93 | Improved |
| use delta-guided bit switching instead of fixed mixed-precision patterns (ablation) | DVD-Quant · Mixed-precision strategies | Imaging Quality: 61.03-61.40 -> 61.93 | Improved |
| increase delta threshold in delta-GBS (ablation) | DVD-Quant · delta threshold sweep | Imaging Quality: avg bit-width decreases, quality gradually drops | Hurt |
| DVD-Quant vs best baseline at W4A4 | DVD-Quant · VBench W4A4 (HunyuanVideo) | Aesthetic Quality: 48.41 -> 61.96 | Improved |
| DVD-Quant W4A6 vs W4A8 baselines | DVD-Quant · VBench W4A6/W4A8 (HunyuanVideo) | Imaging Quality: Imaging 64.47 (SmoothQuant W4A8) vs 64.22 (W4A6) | Mixed |
| combine DVD-Quant with TeaCache | DVD-Quant · Speedup | speedup: up to 4.85x | Improved |
Closest papers by meaning
- Q-VDiT: Towards Accurate Quantization and Distillation of Video-Generation Diffusion Transformers (ICML 2025 · similarity 0.93)
- ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation (ICLR 2025 · similarity 0.92)
- Q-DiT: Accurate Post-Training Quantization for Diffusion Transformers (CVPR 2025 · similarity 0.92)
- VQ4DiT: Efficient Post-Training Vector Quantization for Diffusion Transformers (AAAI 2025 · similarity 0.91)
- $\text{S}^2$Q-VDiT: Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation (NeurIPS 2025 · similarity 0.90)
- PTQ4DiT: Post-training Quantization for Diffusion Transformers (NeurIPS 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
